Using COLMAP Poses and Depth Maps for Dense Reconstruction

[Henry-Bi]

I have a question regarding dense reconstruction. Is it possible to use the poses obtained from COLMAP sparse reconstruction along with measured depth maps to achieve dense reconstruction?

I have a dataset that only provides images and absolute depth maps but does not include camera poses. When visualizing the reconstructed point clouds using trimesh, I noticed that there is always a slight misalignment between the two point clouds.

Do you have any suggestions on how to align the point clouds more accurately or improve the reconstruction quality in this case?

I appreciate your time and look forward to your insights!

[tsattler]

Colmap, as any structure-from-motion software, cannot recover the size of a scene. In fact, the reconstruction can be arbitrarily shifted, rotated, and uniformly scaled without altering the cost function that is optimized by Colmap. Colmap fixes this ambiguity by initializing the coordinate system such that it coincides with the coordinate system of the first image that gets reconstructed such that the second image that gets reconstructed has distance 1 to the first image. The coordinate system then can change during the reconstruction process.

An implication of the above is that the reconstruction is unlikely to fit to the depth maps that you have. You would need to scale it to fit the depth maps. E.g., you can use the sparse points to figure out how to scale the model such that it fits with the depth maps.